Deep learning-based perfusion quantification and large vessel exclusion for renal multi-TI arterial spin labelling...

The multi-TI flow-sensitive alternating inversion recovery sequence is a common ASL technique for probing renal perfusion. However, traditional method for quantifying perfusion, bolus arrival time (BAT) and bolus length (BL) from the images faces challenges due to low signal-to-noise ratio, large vessel contamination, and the absence of magnetization direction information in magnitude images. We proposed a BiLSTM-based deep learning (DL) approach for quantifying perfusion, BAT, and BL, and excluding large vessels. The network was trained on simulated pixel-wise multi-TI signals and tested using simulated and in vivo data. For comparison, the traditional quantification based on Buxton's model fitting was carried out, and manual cortex, medulla, and large vessel masks were drawn on fully relaxed magnitude images. For in vivo data, the quantification results from averages over all repetitions served as reference. In simulation, the DL approach had smaller quantification errors for perfusion and BAT but larger errors for BL than the traditional method. All in vivo parameters derived from the traditional method deviated more from references as number of averages decreased than those derived from DL. The DL masks excluded more high-perfusion pixels than the manual masks. Significant differences between the traditional and DL methods' quantification of in vivo perfusion, BAT, and BL cannot be explained by their differences observed in simulation, suggesting differences between simulated and in vivo data characteristics. The proposed network may serve as a useful tool for quantification in ASL, which is more accurate and more robust against noise than the traditional method.